Abstract
Specialized investment platforms significantly influence financial service operations by enhancing collective intelligence and uncovering trading opportunities. However, the large volume of data, particularly the prevalence of noise, hinders effective decision-making. Opinion leaders constitute a distinct and influential group of users on these platforms, and the question of whether their presence can help financial institutions make better decisions remains open. To address this key operational question, we focus on two indicators in investment platforms: investor disagreement and prediction accuracy. Our findings reveal that the presence of opinion leaders leads to reduced investor disagreement and increased accuracy in predicting future stock returns. This effect is more pronounced when opinion leaders participate earlier in a discussion or contribute posts that are more innovative, longer, or express vivid opinions. Our subsequent experiment further reveals the psychological and cognitive mechanisms behind these findings: the presence of opinion leaders enhances investors’ cognitive abilities by providing them with more information and boosts their confidence in the information they acquire. These results contribute to information cascade theory by demonstrating how varied communication patterns result in different qualities of crowd decision-making. Additionally, we contribute to the operations management literature by highlighting the crucial role of opinion leaders in enhancing both financial services and social media operations. We suggest that financial institutions develop trading strategies based on opinion leaders’ posts and that platforms tailor features to leverage opinion leaders’ participation.
Opinion is the medium between knowledge and ignorance. —Plato
Introduction
With the rise of social media, online forums and discussion boards have become pivotal by enabling investors to exchange information, discuss market trends, and share investment views, for example, StockTwits and Reddit in the United States, and Xueqiu (Snowball) and Eastmoney in China. The collective intelligence generated on these platforms offers valuable signals of market sentiment and is increasingly leveraged by financial institutions and regulators for trading insights and market surveillance.
Among platform participants, opinion leaders can play a pivotal role in shaping investor beliefs and attention on financial social media. Their influence is well illustrated by Keith Gill, known as “Roaring Kitty” or “DeepFuckingValue,” who shared his GameStop investment thesis on Reddit's r/wallstreetbets. His posts attracted widespread attention from retail investors and helped trigger a short squeeze that drove GameStop's share price from below US$20 to over US$500 in early 2021. Given his significant influence, Gill testified before the U.S. House Financial Services Committee that he did not solicit others to trade for personal gain, underscoring his role as an independent opinion leader.
Opinion leaders are often found on social media, where they are typically Internet influencers. In marketing studies, these influencers are shown to affect others’ purchasing behaviors (Kuksov and Liao, 2019). Unlike traditional influencers, who are paid for promotions (Chae et al., 2017) or organizational contexts where employees advocate for company goals (Ballantyne, 2000), the opinion leaders we study operate completely independently of the platform. Their influence is based solely on their demonstrated expertise, as evidenced by their track record of accurate predictions. This independence creates a unique dynamic: while avoiding authenticity issues of sponsored content (Cao and Belo, 2023), they must continually establish their credibility through the quality of their posts rather than institutional affiliation (Enke and Borchers, 2021). How independent opinion leaders influence investor behavior in online investment communities is an important yet underexplored question.
Effective platform operations in online investment communities rely on high-quality information exchange and productive user interactions. For social media operations, a high level of agreement (low level of disagreement) within a stock forum can suggest a harmonious communication environment, fostering a culture that encourages active user participation. The accuracy of predictions made on the platform reflects its professional credibility. When investors consistently make accurate predictions, their satisfaction with the platform increases, which in turn enhances the platform's reputation. Focusing on investor disagreement and prediction accuracy, this paper addresses the following questions:
(1) How does the presence of opinion leaders affect investors’ disagreement and prediction accuracy on financial social media platforms?
(2) What are the mechanisms underlying these effects?
(3) How do communication cues (i.e., opinion leaders’ order of appearance and expression style) moderate these effects?
Using investors’ comments on the same corporate announcement on two stock forums, we apply a difference-in-differences (DID) approach to examine how the presence of opinion leaders affects investors’ belief convergence and decision accuracy. We find that disagreements in the discussion stream decrease when an opinion leader is present, indicating that the opinion leader's presence fosters belief convergence on the platform. Participants’ prediction accuracy improves after an opinion leader joins the discussion, indicating that opinion leaders also help transmit more accurate information and more efficient price discovery. The mechanisms behind these two findings are explored via an experiment. We confirm that the disagreement-reducing effect is due to the image of trust and expertise projected by opinion leaders. Meanwhile, the accuracy-enhancing effect is attributed to the increased cognitive engagement fostered by opinion leaders.
Furthermore, we identify the heterogeneous effects caused by opinion leaders’ order of appearance and style of expression. An opinion leader's early appearance in a discussion accelerates the reduction of disagreements among investors and improves prediction accuracy. Conversely, longer posts by opinion leaders enhance both prediction accuracy and reduce investor disagreement. Additionally, content novelty positively moderates the impact of opinion leaders on investor disagreement, but has no influence on prediction accuracy. Finally, if a post made by opinion leaders includes a clear buy or sell opinion, this will affect investor disagreement and prediction accuracy more significantly than other posts. Specifically, if opinion leaders explicitly talk about bullish or bearish sentiment, compared with opinion leaders who do not express these views, there is a more significant enhancement (approximately 10%) in other investors’ accuracy in predicting stock prices.
Our paper makes several theoretical contributions. First, it advances information cascade theory by showing how opinion leaders accelerate cascades and clarifying the psychological mechanisms underlying communication flows. It further demonstrates how opinion leaders can mitigate reverse cascades and promote collective wisdom, contributing to the debate on whether opinion leaders improve crowd decision quality. Second, we extend the literature on financial service operations. While operations management (OM) research has traditionally focused on portfolio design, fraud detection, and abnormal trading surveillance (Liu et al., 2023; Xu et al., 2017; Zhang et al., 2022), less attention has been paid to authorship signals on social media. Our findings show that opinion leaders in financial forums reduce disagreement and improve prediction accuracy, thereby enhancing the reliability of crowd-based information flows and providing new operational signals for risk management and trading strategies. Third, we contribute to the user-generated content (UGC) platform literature. Prior OM studies largely examine value extraction from UGC (Cui et al., 2018; Ko et al., 2019; Yan et al., 2019; Yan and Pedraza-Martinez, 2019) or structural drivers of engagement such as network size (Wei et al., 2021). We complement this work by showing that opinion leaders act as operational levers shaping content quality and user behavior. Specifically, we demonstrate that the timing, style, and novelty of opinion leaders’ posts moderate their effects on belief convergence and prediction accuracy, offering practical implications for platform operations management.
Literature Review
Social Media in Financial Service Operations
Our study is related to three streams of literature: social media, crowd wisdom, and social media influencers (as shown in Table 1). Social media is an important source of information for financial decision-making. Although most studies focus on sentiment analysis of social media content (Antweiler and Frank, 2004), a few examine factors affecting heterogeneous acceptance of information, such as investor attention and social network effects. For example, Heimer (2016) finds that access to a social network nearly doubles the magnitude of a trader's disposition effect. Schmidt et al. (2020) reveal that Twitter reactions after a glitch accentuate the relationship between supply chain issues and stock market returns. These findings indicate that, in addition to content analysis, social relationships (a core feature of social media) should be integrated into analyses of social media posts to improve financial operations, such as profit optimization or risk avoidance.
Summary of the literature on social media in financial service operations.
Summary of the literature on social media in financial service operations.
In addition to social relationships, the design of social media platforms affects information-based financial decision-making. For example, Corgnet et al. (2024) argue that communication between traders on social media affects market outcomes, especially when the platform publishes a reputation score that can identify a person as untruthful. Cookson et al. (2024) find that attention is strongly correlated across platforms, but sentiment is not. Differences in sentiment across platforms are influenced by variations among users (e.g., professionals vs. novices) and platform design (e.g., character limits in posts).
Crowd wisdom is another area related to stock market prediction. Chen et al. (2014) find that the wisdom of crowds of retail investors can effectively predict future stock returns and earnings surprises. Da and Huang (2020) reveal that herd behavior is more severe when public information includes forecasts from “influential” users. However, they argue that herding with influential users does not improve the accuracy of the final consensus. It should be noted that the opinions in their study are corporate earnings forecasts (including “blind” guesses), while we focus on general opinions based on a particular event. In different scenarios, how and why opinion leaders affect the wisdom of crowds requires further research.
Social media influencers also affect the stock market. For example, Bianchi et al. (2024) examine the information content of tweets from members of the U.S. Congress and their effects on asset prices. They find that tweets from members of Congress contain new and relevant information that is factored into prices over the next few days. It is worth noting that members of Congress differ from common social media influencers, the focus of our study, as they are considered to have access to more insider information.
We identify several research gaps in three areas. In the context of social media, we find that both social networks and platform design have an impact on financial operations, including profit optimization and risk management. This indicates that the role of opinion leaders on social media represents a promising yet underexplored area. Concerning the wisdom of crowds, studies generally treat the crowd as a collective of average investors, but the influence of opinion leaders on crowd behavior warrants further exploration. Additionally, the role of social media influencers, particularly whether retail investors, who are independent in their behavior, can act as influencers, and share insider information within their communities, remains largely unexplored. To address these gaps, exploring how opinion leaders (as influencers) shape the wisdom of crowds on social media emerges as a key research question for both financial and social media operations.
The literature primarily focuses on two established categories of influencers: (1) external influencers who engage in paid promotional activities (Cao and Belo, 2023; Chae et al., 2017), and (2) internal influencers such as employees who advocate based on their organizational affiliation (Ballantyne, 2000; Smith et al., 2021). Although these studies provide valuable insights, they overlook a critical third category emerging on financial social media platforms: independent opinion leaders who operate without formal compensation or institutional ties. Although related constructs, such as organic influencers, unaffiliated experts, or credible posters, have been examined in marketing, communication, and behavioral finance (e.g., Chen et al., 2014; Richins and Root-Shaffer, 1988; Rieh, 2002), we specifically focus on their role in financial information platforms and their impact on the wisdom of crowds and investor behavior, which remains underexplored from an operations perspective.
As Table 2 illustrates, this third category of influencers exhibits fundamentally different characteristics. Unlike paid influencers who may face authenticity challenges, particularly when their promotional nature is not transparent or when they are perceived as large-scale commercial entities (Cao and Belo, 2023), or organizational influencers who struggle with goal alignment (Enke and Borchers, 2021), independent opinion leaders derive their influence solely from their demonstrated expertise. Three key mechanisms drive their influence. First, credibility is earned and sustained through consistent high-quality content and accurate predictions. Second, their persuasive power is strongly correlated with their track record of information accuracy. Third, accountability emerges organically through the collective validation of followers, creating a natural meritocracy where sustained performance determines influence.
Comparison of influencer types.
Comparison of influencer types.
Information Cascade Theory
In many scenarios, decision-making is a collective process in which multiple people make decisions in sequence, leading to information cascades (Tump et al., 2020). This phenomenon, whereby early decisions influence subsequent choices, involves two steps. Initially, an individual faces a situation that requires decision-making. Subsequent decisions are made sequentially, with each person observing the choices of those before them. An information cascade occurs when the influence of others’ decisions supersedes personal information (Duan et al., 2009), leading to a scenario where “an individual's action does not depend on his private information signal” (Bikhchandani et al., 1992: 1000). In such instances, the decision-maker tends to follow their predecessors’ decisions, disregarding their private information.
Information cascades have been intensively studied in the OM literature. For example, herd behavior of joining a long queue is associated with ex-post regret, which affects customer experiences (Veeraraghavan and Debo, 2011). Using the information cascade, a wise decision-maker can infer others’ private information using response times (Frydman and Krajbich, 2022). Glazer et al. (2021) focus on the sequential investment process and find that the occurrence of information cascades depends on the cost of acquiring information. In the context of crowdfunding, information cascades are found in herd transactions. Xiao et al. (2021) reveal that the frequency of communication messages has a positive effect on backer contributions, but attenuates successors’ herd momentum toward their predecessors. The closest paper to ours is Ke et al. (2024) who examine the factors that affect the likelihood of word-of-mouth cascades on social networks. The authors find that cascading between a followee and a follower is less likely when the followee is a high-status member, a woman, or has a strong connection to the follower, and more likely when the product considered is inexpensive. In summary, the OM literature explores the antecedents and consequences of information cascades, revealing their managerial implications for crowdfunding, marketing, and customer experiences.
This paper contributes to information cascade theory in three ways. First, although the two-step process of information cascades, that is, individuals making decisions based on their private information initially and later following others’ choices when observing their actions, is commonly observed on social media platforms between followers and followees (Lee et al., 2015), there is little research on whether and why information cascades occur between social media influencers and unconnected users, especially in the context of financial investors. If such cascades occur, information may bypass incremental diffusion through direct connections and instead spread rapidly through the broader network, supported by underlying psychological mechanisms.
Another debate about information cascades is whether they lead to wise or mad crowd opinions (Tump et al., 2020). Understanding the conditions that lead to positive and negative information cascades is crucial for many fields. We focus on the financial market and investigate information cascades from opinion leaders, who are the wisest in the crowd, to the general public. Surowiecki (2005) posits that groups can be incredibly intelligent, often surpassing the smartest individuals within them. This suggests that information cascades may result in suboptimal decision-making. We contribute to this ongoing debate by empirically investigating whether information cascades have positive or negative effects on collective wisdom.
Third, information cascades are often influenced by information cues, but the literature primarily focuses on moderators such as product features and seller characteristics (e.g., Tan et al., 2019). By considering communication features such as order of appearance and expression style, we seek to uncover the moderating roles of these signals in influencing information seekers’ rational decision-making. Our findings enrich information cascade theory by elucidating the specific conditions that lead to diverse information cascade outcomes.
Research Model
We present our research model in Figure 1. The information cascade occurs between an opinion leader and their followers. We focus on two crowd-level measures, disagreement and accuracy, which are important operational indicators for both financial institutions and social media platforms. Disagreements among investors suggest a dispersion of beliefs among investors, while prediction accuracy indicates the potential for efficient price discovery. These indicators are not only essential trading signals that enable financial institutions to develop effective strategies to make profits but also key factors in the operation of financial service platforms. Agreement among investors suggests a harmonious communication environment, while disagreement harms user participation. Accuracy reflects the effective price discovery and professional level of a financial service platform. When investors make correct predictions on the platform, they are likely to be satisfied with the platform, thus enhancing its reputation and attracting more users and opinion leaders. The presence of more professional participants can, in turn, further increase prediction accuracy.
Research model.
Drawing on information cascade theory, we hypothesize that the presence of opinion leaders will reduce investor disagreement (H1) and increase prediction accuracy (H2). The order of appearance and style of expression are key communication features that affect investors’ perception of information. We thus include these features as moderators and propose H3a, H3b, H4a, and H4b accordingly.
On social media, information often flows in cascades across a social network. Various structural and temporal features of a network influence the virality of cascades and lead to diverse outcomes. In this study, we focus on opinion leaders in social networks. Opinion leaders are recognized as having considerable influence on community opinions and trends. These leaders, often social media influencers or Internet celebrities, have large followings and exert influence through their popularity and expertise (Li, 2018; Rapanos, 2023).
As indicated by Shiller (1995), an information cascade is a rational information-based interpretation process, in which individuals rationally ignore their own signals and decide to follow others. Many experiments illustrate that information effects, not social effects, are the main factor in information cascades. Based on this perspective, we divide the identity of opinion leaders into two important information signals: trust and expertise.
First, opinion leaders create a sense of trust among their audience by investing considerable effort in image building (Kim and Kim, 2021). Huang (2015) notes that the more popular a social media influencer is, the more trustworthy they are perceived to be. This phenomenon is supported by the trust transfer literature, which posits that people's trust in influencers extends to the influencers’ posts. For instance, Hermanda et al. (2019) find that recommendations from social media influencers are more trusted than those from family or friends.
Second, the expertise of opinion leaders accelerates the cascade of information. “Expert power,” which denotes the persuasive influence of a highly knowledgeable individual, is well documented in the communication literature (Klucharev et al., 2008). In our context, opinion leaders are identified as individuals who have been active on stock forums for an extended period and have published a large number of posts. This experience is considered a signal of expertise. The label of expertise of opinion leaders will increase their persuasiveness and facilitate the cascade of information among information seekers. In addition, opinion leaders mainly rely on the quality of their posts to attract followers (Li et al., 2011). High-quality posts convey cognitive authority, enhancing opinion leaders’ persuasiveness (Rieh, 2002), which further enhances the cascade of information.
Regarding financial markets, Hong and Stein (2007) find that investor disagreements around public announcements often stem from different interpretations of these announcements. Our study predicts that opinion leaders on social media, due to their trusted and authoritative status, play a pivotal role in interpreting and disseminating firm announcements. Their interpretations, stemming from their image of trust and expertise, are likely to be adopted by their followers (Li et al., 2022), leading to increased investor confidence in these interpretations. Consequently, this heightened confidence in the information provided by opinion leaders is expected to facilitate the convergence of investor beliefs, thereby reducing disagreements among investors.
H1. (Disagreement-Reducing Effect). The presence of opinion leaders in a discussion stream increases investors’ confidence in their information (through opinion leaders’ image of trust and expertise), thus reducing disagreements among investors.
Accuracy-Enhancing Effect
Information cascades result from rational inferences that others’ decisions are based on information that dominates private signals. Of particular interest is the potential for a reverse cascade, in which early decision-makers unfortunately encounter private signals suggesting an incorrect outcome and followers subsequently adopt this erroneous pattern (Anderson and Holt, 1997).
The influence of opinion leaders on information cascades affects cascade outcomes in two ways. First, opinion leaders are often domain experts who tend to make extensive use of their private signals (Alevy et al., 2007). Their likelihood of making mistakes is much lower than that of amateurs. As noted by Lou (2022), opinion leaders consistently create and share content that is both useful and deeply rooted in their specific knowledge domain. This activity underscores their expertise, a concept that Richins and Root-Shaffer (1988) associate with an inherent “knowledgeable feature” of opinion leaders. The concept of opinion leaders’ “knowledge ability,” as examined by Myers and Robertson (1972), reveals a significant correlation between opinion leadership and various measures of knowledge. This finding is crucial, as it implies that the depth and breadth of an opinion leader's domain knowledge translate directly into the accuracy and reliability of the information they provide. The more knowledgeable an opinion leader is, the more precise and actionable their insights become. Furthermore, Goes et al. (2014) point out that opinion leaders with large followings tend to communicate in a manner consistent with their “expert” status. This behavior involves not only the dissemination of information but also the collection and synthesis of insider information and innovative ideas. Such efforts contribute significantly to maintaining their status as experts and, more importantly, to enhancing the quality of the information they share.
Second, opinion leaders will enhance the cognitive level of participants. Indeed, by engaging with content from these knowledgeable and expert opinion leaders, investors are exposed to high-quality and insightful analyses and interpretations of market trends and dynamics. This exposure boosts their cognitive understanding of the market, enabling them to process, analyze, and interpret market information more effectively and accurately. Furthermore, the presence of opinion leaders can lead participants to perceive the topic as important, prompting them to pay greater attention to collecting relevant information. In such cases, the cascade might shift to a correct state, a phenomenon known as “self-correction” (Goeree et al., 2007).
Therefore, the presence of opinion leaders on financial social media platforms plays a pivotal role in increasing the cognitive level of investors. This increased cognitive capacity, fueled by expert insights and high-quality information, is a key factor in improving the accuracy of investors’ predictions of future stock returns. The role of opinion leaders thus extends beyond information sharing to significantly influence the cognitive processes and decision-making abilities of investors in the financial market.
H2. (Accuracy-Enhancing Effect). The presence of opinion leaders increases investors’ cognitive level, thus increasing the accuracy of their predictions regarding future stock returns.
Moderators
Order of Appearance
Information cascades often occur in a sequential model, in which agents take actions sequentially after observing action history and a private signal (Lee, 1993). The order in which opinion leaders appear in a sequential model influences the outcomes of information cascades. If an opinion leader appears early in the sequence, information seekers are likely to mimic their actions, which subsequent participants can then observe. According to the majority rule in information cascades (Hung and Plott, 2001), more consistent actions among participants lead to greater convergence. Therefore, an opinion leader who appears earlier in the sequence will have a greater impact on the information cascades than one who appears later, due to the majority effect triggered by their early presence.
H3a. The order of appearance of opinion leaders negatively moderates the disagreement-reducing effect.
As mentioned earlier, accuracy is enhanced through better information quality and increased cognitive engagement. In a sequential model, the timing of an opinion leader's appearance has a significant impact. In information cascades, individuals make decisions by gradually gathering and processing information until a certain threshold is reached, leading to a decision. The internal cognitive process of accumulating evidence over time is called “the information accumulation effect.” As more information accumulates, an opinion leader's statement becomes more informative and targetable. Therefore, an opinion leader who appears later in the sequence tends to have a more pronounced impact on enhancing prediction accuracy, which leads to H3b:
H3b. The order of appearance of opinion leaders positively moderates the accuracy-enhancing effect.
Expression Style
Drawing on information cascade theory, we identify two key informative signals—trust and expertise—that influence investors’ confidence in the information provided by opinion leaders. These two signals are strongly related to the content provided by opinion leaders. First, high-quality information is essential to establish an image of trust and expertise. Following the literature, we measure post quality based on linguistic features, including word count and text newness (Shin et al., 2020). Text length is often considered a reliable indicator of quality, as longer texts typically contain more information and evidence. Text newness quantifies the distinctiveness of an individual post compared with the typical or average content of posts. Higher-quality information (i.e., longer and more recent information) provided by opinion leaders will enhance their image of trust and expertise, which will further improve information seekers’ confidence in opinion leaders, thus accelerating the cascade of information.
Another important signal for expertise is the polarity of opinions. According to Lewin Loyd et al. (2010), when they express their expert knowledge using a more powerful style of speech, experts are more liked, more influential, and generate more confidence. In our context, a powerful style is determined by whether a post clearly expresses buy or sell suggestions, a concept we refer to as “opinion polarity.” We propose that clear opinion polarity in an opinion leader's post will accelerate the cascade of information.
H4a. A powerful expression style (i.e., longer text, more recent content, and straightforward opinion) positively moderates the disagreement-reducing effect.
Opinion leaders are considered to influence investors’ prediction accuracy by enhancing their cognitive ability. Longer text and more recent content provide more information for decision-making. Longer texts often contain more detailed explanations and data. This depth allows decision-makers to have a clearer understanding of the current situation. New posts introduce new perspectives or unique analyses that have not been considered before. This can lead investors to rethink their strategies or consider factors that they previously overlooked, potentially leading to more accurate market predictions. Regarding posts with clear opinions, it is common for a post to present detailed arguments before reaching a definitive buy or sell conclusion. Such straightforward opinions are easier to follow and reduce the risk of misunderstandings. Therefore, we propose that clear opinion polarity in an opinion leader's post will enhance prediction accuracy.
H4b. A powerful expression style (i.e., longer text, more recent content, and straightforward opinion) positively moderates the accuracy-enhancing effect.
Data and Variables
Stock Forum Data
Our data come from two main stock forums in China: Eastmoney (www.eastmoney.com) and Snowball (https://xueqiu.com). These are the two largest and most popular Chinese stock forums, where investors share their opinions on stocks and stock announcements. They bring together a considerable number of stock investors who post their beliefs and reveal their trading strategies. Both websites have a bulletin board system-like format, providing online platforms where investors can exchange information in real time. The two platforms cover all announcements of Chinese listed companies (3,557 in 2019) and allow users to post comments under each stock announcement. At the end of 2018, Eastmoney and Snowball had 33.78 million and 4.22 million monthly active users, respectively, accounting for 26% of the total number of individual A-share investors in China. Therefore, it follows that user data from these two platforms are sufficiently representative of the Chinese stock market. 1
Distribution of the number of comments.
Frequency of positing comments and announcements.
We obtain stock announcement information, related comments, and the user profile information of comment authors from the two platforms for the period from January 2019 to December 2019. 2 Our dataset covers 296,511 announcements 3 published on the two stock forums, for a total of 3,557 stocks. Each downloaded stock announcement has a unique identifier consisting of the company's stock ticker symbol, announcement date, announcement time, and sequence number. Each comment identifier contains the stock announcement ID, comment date, comment time, and a unique comment ID. Figure 2 shows that for our sample, the minimum number of comments posted under each announcement is 0 and the maximum number is 1,095. Approximately 75% of the announcements have more than six but fewer than 21 comments on both platforms.
Figure 3 plots the hourly distribution of comments on the two platforms. We observe a different hourly distribution of online postings than Antweiler and Frank (2004), who find that the entire message board they study is quiet between midnight and 07:00 and that message traffic is highest during trading hours, staying active until the evening. However, in our sample, we find that most comments are made after the market closes, between 15:00 and 23:00 (panel A of Figure 3). One possible explanation for this phenomenon is that Antweiler and Frank (2004) focus on investors’ opinions regarding stock performance, and most discussions about stock performance take place during trading hours. In contrast, our data focus on investors’ opinions regarding stock announcements, and most Chinese stock announcements are issued after the market closes (panel B of Figure 3). Therefore, the most active time for investors to discuss their opinions on stock announcements is after the market closes. From Figure 3, we observe that the temporal distribution of online posts is consistent with the publication time of stock announcements, and nearly 80% of the announcements in our sample are made after trading hours. This finding supports the conclusion that investors update their messages in real time as financial events unfold.
One of the challenges we face is defining an opinion leader. Communication theory asserts that an opinion leader is a respected, trusted, and active media user with experience and expertise in a particular field. An opinion leader interprets the meaning of media messages or content for lower-end media users. Following the above definition, in our sample we define opinion leaders on the two stock forums according to their number of followers. Given the characteristics of the stock forum as a community for sharing investment knowledge, the number of followers indicates the user's influence on other users in the community. We identify the top 5% of users on Eastmoney.com and Snowball.com, in terms of number of followers, as opinion leaders; that is, we capture the top 5% of media users as our opinion leaders. 4 For prudential considerations, we additionally use the top 2% and 10% of users on the two stock forums as the cutoff in robustness checks, to address the problem of some specific investors acting as an omitted variable caused by the 5% selection.
Measure of Investor Disagreement
We directly measure investor sentiment using the content of each investor post under stock announcements on the two stock forums using the sentiment analysis technique. One popular method in the literature for analyzing sentiment is to count the number of negative words in a passage (Chen et al., 2014; Tetlock et al., 2008). However, developing a sentiment dictionary for the financial domain in Chinese presents significant challenges. To solve this problem, we apply the advanced supervised machine learning algorithm BERT to classify each comment's content as positive, negative, or neutral, and then generate the investor sentiment index. Online E-Companion A explains the BERT method, compares it with other popular algorithms, and proves its efficiency.
The BERT algorithm helps us to first obtain a sentiment index, sentimentij, which measures the general opinion of online comment j regarding stock announcement i; it takes the value of {−1, 0, 1}, corresponding to pessimistic, neutral, or optimistic attitudes toward the stock announcement. Then, we construct the investor disagreement index following Antweiler and Frank (2004), using the following equation:
Constructing the investor disagreement index requires adjustments to our analysis. One source of potential concern is how to treat periods during which no investor comments are posted. In our analysis, we assume that investors are neutral in a normal state and that changes in sentiment are caused by the publication of announcements and reflected in posted comments. Therefore, in the reported results, we assume that periods without comments correspond to periods with zero investor disagreement. Second, when we match the comment message data (used to contrast the investor disagreement index) with the trading data, we assign all messages posted after 15:00 on trading day t − 1 and before 09:30 on trading day t to trading day t.
We construct a proxy to measure the prediction accuracy of investor comments in a manner similar to Chen and Hwang (2022). We evaluate the accuracy of stock predictions made by investors who comment on stock announcements as the alignment of sentiment expressed by other commentators in response to the stock's abnormal return movement in the market over the next 3 trading days. A stock's abnormal returns are the differences between its actual and predicted returns, and we use three-factor models to compute expected returns: the capital asset pricing model (CAPM), the Fama–French (1993) three-factor model, and the Fama–French (2015) five-factor model.
5
To measure a stock's predicted returns, we first run the following regression model to estimate the factor loadings (β1, β2, and β3) for the three factors in the Fama–French three-factor model using historical stock returns:
We then calculate the predicted return on the asset using the estimated factor loadings and the actual factor values for the period we are analyzing using the following equation:
After estimating the predicted return from the three-factor models, we calculate the abnormal return as the difference between the actual return and the predicted return. Abnormal returns are typically used in finance to measure the performance of a stock or portfolio relative to a benchmark or market index.
Finally, we construct our measure of the prediction accuracy of investor comments (Accuracy). Accuracy takes a value of 1 if either a comment's sentiment/tone is above 0 and the corresponding stock generates positive abnormal returns over the next 3 trading days or a comment's sentiment/tone is below 0 and the corresponding stock generates negative abnormal returns over the next 3 trading days; otherwise, Accuracy takes a value of 0. For robustness tests, we also construct an analogous measure of the accuracy of opinion leaders’ comments based on abnormal returns over the next 5 trading days.
Identification Strategy
We use the DID estimation method to control for endogeneity issues. Eastmoney.com and Snowball.com are the largest Chinese social media platforms for sharing investment ideas and strategies, and both forward published stock announcements to their discussion boards in real time. Therefore, we can simultaneously observe the two forums’ discussion threads concerning the same announcement on different websites. We match the two comment threads for the same announcement on the two websites as a comment–thread pair. Next, we define the discussion thread type: threads with opinion leaders constitute the treatment group and threads without opinion leaders constitute the control group. Accordingly, both groups consist of announcements from Eastmoney.com and Snowball.com, and whether a thread is assigned to the treatment group depends on whether opinion leaders participate in the discussion. Finally, we choose thread pairs—one from the treatment group and one from the control group for the same firm announcement—to construct our DID estimation sample. An example of our identification strategy is provided in online E-Companion B.
Endogeneity Concerns. Our DID setting addresses endogeneity problems in several ways. First, our DID sample consists of discussion threads listed on two stock forums, and the condition for being assigned to the treatment group is the thread where the opinion leaders appear. Therefore, our treatment and control groups are composed of discussion threads from both Eastmoney.com and Snowball.com, eliminating the impact of different investor bases on the two platforms. Second, we pair two discussion threads about the same announcement on both forums, and the two discussion threads start at the same time. Using the difference between the treatment and control discussion threads in a pair helps eliminate the impact of stock price changes, announcement topic, stock characteristics, and time series. Third, because each comment in discussion threads has a timestamp, we can observe the time of appearance of opinion leaders and measure changes in investor disagreement before and after the appearance of opinion leaders. Using the difference before and after the appearance of opinion leaders helps us eliminate the impact of stock forum and discussion thread characteristics. This setting enables us to shed light on the direction of causality between the presence of opinion leaders and changes in investor disagreement, thus avoiding reverse causality concerns.
Eliminate Followers and Institutional Investors. One concern is that the reduction in participant disagreement may be caused by the “follower effect.” Given that opinion leaders are defined as users with many followers, when opinion leaders post comments, they are likely to trigger comments from their followers. If this is the case, it is not surprising that we observe a decrease in disagreements after the appearance of opinion leaders. To eliminate this issue, we collect the names of followers of each opinion leader in our sample and eliminate their posts from the comment threads. In this way, we estimate only changes in the disagreement level of nonfollowers before and after the appearance of opinion leaders. Likewise, the estimation results should not stem from followers being triggered by a post from an opinion leader.
Another concern is that our results may come from institutional investors. Institutional investors may attract more followers, thus making it easier for them to be defined as opinion leaders. In this case, our results would come from the difference between institutional investors and retail investors. Thus, we exclude institutional investors 6 from our empirical sample and find that our empirical results remain robust.
Parallel Trend Test. We conduct a parallel trend test for our DID sample in Figure 4(a) and (b) to test whether the treatment and control groups in our DID sample exhibit similar disagreement levels and accuracy levels before the appearance of opinion leaders. We thus consider a 13-event window comprising six comments before a post by the opinion leader and seven comments after it. The dashed lines represent 95% confidence intervals (CIs), adjusted for announcement-level clustering. Figure 4(a) demonstrates that the difference in the level of investor disagreement between the treatment and control groups before the appearance of opinion leaders is not statistically significant. However, after the appearance of opinion leaders, this difference becomes statistically significant and negative. This indicates that, compared with the control group, disagreement in the treatment group decreases significantly. Figure 4(b) presents the parallel trend test for accuracy. We observe that before the appearance of opinion leaders, there is no statistically significant difference in investors’ accuracy levels between the treatment and control groups. However, following the appearance of opinion leaders, there is a statistically significant and positive difference in investor accuracy between the treatment and control groups. This suggests that, compared with the control group, the treatment group experiences a substantial increase in investors’ prediction accuracy.
Parallel trend tests.
Summary Statistics of the DID Sample. In panel A of Table 3, we show the summary statistics of the stock forum data in our DID sample. Column 1 describes the data for the full sample, column 2 describes the data from Eastmoney.com, and column 3 describes the data from Snowball.com. We can see that the average investor sentiment and average opinion leader sentiment of the two platforms are very similar and are both around 0. This finding supports the assumption that there is no difference between the average sentiment of investors on the two platforms. Given that we mainly compare changes in investor sentiment between the treatment and control groups, it is reasonable to combine the data from the two platforms to construct our DID sample. We show the summary statistics for threads with and without opinion leaders in panel B of Table 3. Based on the summary statistics, we observe that threads without opinion leaders have an average sentiment of −0.031, while threads with opinion leaders show an average sentiment of −0.022 before their appearance and −0.031 after. Importantly, the average sentiment remains relatively consistent across all three comment groups, hovering around 0. This suggests a similar sentiment orientation in the comments, with no substantial differences between threads without opinion leaders, threads with opinion leaders before their appearance, and threads with opinion leaders after their appearance. Additionally, there is a relatively small disparity in investor disagreement between threads without opinion leaders and those with opinion leaders before their appearance, indicating a comparable level of investor disagreement in both groups before the introduction of opinion leaders. Moreover, posters in threads without opinion leaders and threads with opinion leaders before their appearance show a minor difference in prediction accuracy. This suggests a similar level of accuracy in predicting investor behavior between these two groups prior to the arrival of opinion leaders. In contrast, the treatment group exhibits a decrease in investor disagreement from 0.655 to 0.345, and an increase in investor prediction accuracy from 0.260 to 0.295 after the appearance of opinion leaders. This finding aligns with our main empirical results.
Summary statistics of stock forum data—DID sample.
Note. DID = difference-in-differences.
Disagreement-Reducing Effect
After constructing our DID sample, we study whether the appearance of opinion leaders causes the convergence of other participants’ beliefs. We estimate the impact of opinion leaders on investor disagreement using the following regression equation:
Table 4 reports the results of equation (5). Columns 1–4 present four econometric models with different fixed effects to avoid potential endogeneity problems caused by omitted variables. We find that in all econometric models, the interaction coefficient between the treatment group and the After dummy is negative and significant at the 1% level. For example, the regression coefficient on the interaction coefficient between Treat and After is −0.0510 when we control for comment time trend, trading day, and platform fixed effects. This finding shows that on average, discussion threads with opinion leaders exhibit a 5.1% decrease in disagreement after the appearance of opinion leaders. This result is consistent with H1, which posits that the appearance of opinion leaders decreases disagreement over a firm announcement.
Opinion leaders and investor disagreement.
Notes. The t-statistics are reported in parenthesis.
Significance levels: *p < 0.1, **p < 0.05, ***p < 0.01.
For this specification, we use comment accuracy as a dependent variable in the regression and are interested in the interaction coefficient on
We report our results in Table 5. We consider three measures of comment accuracy:
Opinion leaders and prediction accuracy.
Opinion leaders and prediction accuracy.
Notes. The t-statistics are reported in parenthesis. CAPM = capital asset pricing model; FF3 = Fama–French three-factor model; FF5 = Fama–French five-factor model.
Significance levels: *p < 0.1, **p < 0.05, ***p < 0.01.
As the variable After in Section 5.2.2 is defined as the period following the appearance of the first opinion leader in the treatment group, it fails to account for the impact of an additional opinion leader on other investors’ disagreement level and prediction accuracy. In this section, we investigate the incremental impact of each additional opinion leader as they emerge. This marginal effect can help us to better understand the dynamic influence of opinion leaders. To explore this impact, we use a continuous treatment model, specifically adopting a dose–response function (DRF) approach adopted in Cerulli (2014). In this model, we use (a) the number of appearances of opinion leaders and (2) the aggregate influence of opinion leaders as measures of treatment intensity. Subsequently, we use the DRF approach to reflect variations in the impact of the number and overall sentiment of opinion leaders on investors’ disagreement and accuracy. Online E-Companion C illustrates the DRF results.
Robustness Checks
To establish the internal validity of our results, we conduct a battery of robustness checks (refer to online E-Companions D–I for details). In online E-Companion D, we use the coarsened exact matching method as a robustness check. Using average comment sentiment, number of comments, and average comment length as treatment variables, we conduct coarsened exact matching to refine our model and rerun our analysis. In online E-Companion E, we keep only commenters who post both before and after the emergence of opinion leaders to test whether our main results still hold when we eliminate the difference in the investor base. In online E-Companion F, following Baker et al. (2022), we include never-treated units (discussion threads without opinion leaders that cannot be matched to a treatment thread) in our staggered DID estimation sample as clean controls. In online E-Companion G, we narrow and expand the definition of opinion leaders by considering users with the most followers, ranging from the top 2% to the top 10% on both platforms. In online E-Companion H, we replace the disagreement index with linear or entropy functions. In online E-Companion I, we extend our analysis by using time-lagged models to investigate whether the influence of opinion leaders persists beyond the immediate treatment period. We find that the disagreement-reducing effect disappears in comments after 14 days. Conversely, the accuracy-enhancing effect persists and even strengthens 1 to 2 weeks after the appearance of opinion leaders.
Moderation Effects of Order of Appearance and Expression Style
Order of Appearance of Opinion Leaders
In this section, we construct a variable, appearance order, representing the sequence number of the first opinion leader's comment in the comment stream. For instance, if the first opinion leader's comment appears as the third comment in the stream, appearance order is set to 3. Therefore, a higher value of appearance order indicates that the opinion leader's comment appears later in the sequence of comments. Table J1 in online E-Companion J illustrates the impact of the position of opinion leaders’ comments on investor disagreement. We observe that the interaction term between appearance order and After on investor disagreement is negative and significant at the 1% level. This finding suggests that when the first opinion leader appears later in the sequence of comments, the reduction in investor disagreement becomes more pronounced. One possible explanation is that in the information cascade process, opinion leaders who appear later may gather insights from previous comments and formulate more in-depth and compelling views based on their accumulated knowledge. This process of information accumulation can enhance the persuasiveness of their comments, thereby increasing their ability to influence others (Bikhchandani et al., 1992). H3a is therefore supported. The order of appearance of opinion leaders actually weakens, or negatively moderates, the disagreement-reducing effect. Table J2 in online E-Companion J presents the impact of the position of the first opinion leader's comments in the comment stream on investor accuracy. The results indicate that when the first opinion leader appears later in the comment stream, there is a more pronounced enhancement in the overall accuracy of predictions regarding future stock prices. In other words, when opinion leaders appear later in the sequence, other investors tend to make more accurate predictions. These results confirm H3b.
Expression Style
Another interesting empirical question is whether the effects of opinion leaders on the convergence of other participants’ opinions vary depending on what the opinion leaders talk about. Therefore, we conduct a content analysis for the case of a single opinion leader. We analyze three characteristics of content: length, newness, and sentiment expressed.
In panel A of Table K1 in the online E-Companion, we interact comment length with the interaction term between Treat and After. The length of opinion leaders’ comments is calculated using the number of words. We find that the longer the comments made by opinion leaders, the greater the decrease in other participants’ disagreement in the treatment group after the appearance of opinion leaders. Given that the literature shows that word count or content length is positively correlated with information quality (Chen and Hwang, 2022), our findings suggest that the more information an opinion leader provides, the more belief convergence occurs in the discussion thread. Similar results are obtained in panel A of Table K2, particularly in terms of the impact of comment length on investors’ prediction accuracy. We find that the longer the comments of opinion leaders, indicating better information quality, the more significant the enhancement in investors’ accuracy in predicting stock prices.
In panel B of Table K1 and panel B of Table K2, we present the impact of the newness of the information conveyed by opinion leaders on disagreement and accuracy, respectively. Newness is measured by whether a comment is distinct from previous comments on the same announcement. We first calculate the cosine similarity of a comment with every comment posted before it and take the maximum value. Next, we use 1 minus the maximum cosine similarity calculated in the previous step to represent the newness of a comment. The newness index ranges from 0 to 1, with smaller values indicating more homogeneous content than larger values. Table K1 shows that the triple interaction coefficient between opinion leaders’ comment newness, Treat, and After is negative and significant at the 1% level, suggesting that the newness of opinion leaders’ comments is correlated with a greater decrease in investor disagreement in the treatment group after the appearance of opinion leaders. This finding suggests that investor disagreement decreases more when opinion leaders contribute more new information than when they contribute less new information. However, we do not observe a systematic effect of the newness of opinion leaders’ comments on the accuracy of stock price predictions by other investors. One possible explanation is that when other investors pay attention to the market forecasts made by opinion leaders, they are primarily interested in the direction of the market trend revealed by the opinion leaders—whether it predicts an increase or decrease. The focus may not be on whether opinion leaders’ comments contribute sufficient new information.
Finally, we construct the dummy variable opinion polarity to indicate whether the opinion leader explicitly indicates whether the stock price will go up or down. If an opinion leader's comment explicitly expresses an opinion regarding the rise or fall of a stock, the variable is set to 1, and otherwise 0. In panel C of Table K1 and panel C of Table K2, we find that when opinion leaders’ comments explicitly predict whether the stock will rise or fall, the reduction in disagreement among other investors and the improvement in the accuracy of their predictions regarding future stock prices become more pronounced. H4a and H4b are therefore supported.
Experiment
Our empirical analyses reveal that when an opinion leader appears in a discussion stream, investors’ disagreement decreases and their prediction accuracy about future stock returns increases. However, some unobservable factors (e.g., factors inferable from the text of posts in the thread, forum features) may bias our results and hinder a causal interpretation. To resolve this issue, we conduct a controlled experiment to further support the causality interpretation.
Participants and Procedure
This experiment involved a two-way (forum type: with opinion leader vs. without opinion leader) between-subjects design. A total of 200 participants from Credamo (61.0% female; Mage = 30.72, SDage = 7.37) completed all of the tasks in this experiment. The participants were randomly assigned to one of the two scenarios. All information in these two scenarios was identical, except for the labeling of one participant as an opinion leader in the “with opinion leader” condition.
At the beginning of the experiment, the participants were asked to imagine browsing a stock's overview on an online forum, where they were shown the stock's actual semiannual report and investor comments published after the market closed on Friday. For brevity, we selected six comments from the forum. In the “with opinion leader” condition, one comment (“It will drop by 3% next Monday”) was marked as coming from the opinion leader. In contrast, the “without opinion leader” condition featured no marked opinion leader.
To measure the level of disagreement, we asked the participants: “After reading these comments, are you more inclined to think that this stock will go up or down?” (1 = Rise, 2 = Unsure, 3 = Fall). To assess prediction accuracy, the participants were asked to predict the stock price on Monday (ranging from −10.00% to 10.00%). The participants’ perceived trust in the forum's information was measured using three items rated on a 7-point scale (e.g., “I find the information provided by forum participants to be professional and credible”; 1 = strongly disagree, 7 = strongly agree; α = .93), adapted from prior research (McKnight et al., 2002; Pavlou and Gefen, 2004). Finally, the participants’ perceived financial cognitive level was measured using three items rated on a 7-point scale (e.g., “I am able to predict market trends more accurately because of the information I obtained from this forum”; 1 = strongly disagree, 7 = strongly agree; α = .90), adapted from Huston (2010).
Results
Level of Disagreement. A chi-square analysis revealed a significant difference in the level of disagreement between conditions (χ2(2) = 17.79, p < 0.001). As predicted, a higher proportion of participants in the “with opinion leader” condition (76 out of 100, or 76%) predicted a stock price fall relative to the “without opinion leader” condition (53 out of 100, or 53%).
Prediction Accuracy. The actual stock price fell by −2.88% on Monday. We calculated the absolute value of the difference between the participants’ predictions and the actual fall, with a higher value indicating lower accuracy and a lower value indicating higher accuracy. As anticipated, the presence of an opinion leader significantly affected the participants’ prediction accuracy (F(1, 198) = 6.03, p = 0.015), with higher accuracy being observed in the “with opinion leader” condition (M = 2.64, SD = 2.81) than in the “without opinion leader” condition (M = 3.55, SD = 2.41).
Mediating Role of Perceived Trust in the Level of Disagreement. In line with our hypothesis, the participants in the “with opinion leader” condition demonstrated higher perceived trust in the information disclosed on the forum (M = 5.03, SD = 1.45) than those in the “without opinion leader” condition (M = 4.42, SD = 1.62; F(1, 198) = 7.79, p = 0.006). To test our hypothesized mediation model (i.e., whether the effect of opinion leader presence on disagreement is transmitted through perceived trust), we conducted a mediation analysis using PROCESS macro for SPSS, Model 4 (Hayes, 2017). This model employs an ordinary least squares regression-based path analysis and a bootstrapping procedure (with 5,000 resamples) to estimate the indirect effect. Our mediation analyses confirmed that the appearance of an opinion leader influenced the level of disagreement through perceived trust (B = 0.17; SE = 0.06; 95% CI: .05 to .30).
Mediating Role of Perceived Financial Cognitive Level in Prediction Accuracy. Consistent with our prediction, the participants in the “with opinion leader” condition exhibited a higher perceived financial cognitive level (M = 5.41, SD = 1.09) than those in the “without opinion leader” condition (M = 4.48, SD = 1.51; F(1, 198) = 43.87, p < 0.001). Furthermore, our mediation analyses (PROCESS Model 4 with 5,000 bootstrap samples; Hayes, 2017) confirmed that the effect of the presence of an opinion leader on prediction accuracy was mediated by perceived financial cognitive level (B = 0.74; SE = 0.27; 95% CI: .30 to 1.35).
Discussion. The results of this experiment support our hypothesis that the presence of opinion leaders enhances investors’ confidence in the information disclosed by the company, thereby reducing disagreements among investors, and increases their cognitive level, thereby improving their prediction accuracy regarding future stock returns. The controlled experiment rules out potential alternative explanations and provides causal evidence for our findings.
Discussion and Conclusions
Discussion of Key Findings
Our findings demonstrate that opinion leaders play a central role in shaping both financial service operations and social media dynamics by reducing disagreement and enhancing prediction accuracy. The observed convergence in beliefs indicates that opinion leaders reduce the noise created by dispersed and conflicting opinions, thereby streamlining the flow of information. This aligns with research on information cascades and social influence, which highlights the efficiency gains of coordinated attention in noisy environments (e.g., Ke et al., 2024; Xiao et al., 2021). At the same time, the improvement in prediction accuracy suggests that opinion leaders are not merely amplifying herd behaviors but are also synthesizing and conveying information in ways that raise the overall informational quality of the forum. This addresses prior concerns that cascades suppress individualism (Wang et al., 2018), showing instead that collective efficiency can benefit when high-quality opinion leaders guide discussions.
We also identify heterogeneity in the mechanisms of influence. Early intervention by opinion leaders accelerates consensus, while expressive strategies such as longer and more elaborate posts enhance both accuracy and agreement, consistent with impression management theories (Chen and Hwang, 2022). Interestingly, content novelty decreases disagreement but does not improve accuracy, suggesting that while novelty may capture attention and foster alignment, it does not always guarantee information accuracy. Likewise, posts with explicit buy or sell recommendations have a disproportionately strong effect, increasing prediction accuracy by about 10%. This finding is consistent with the literature on the persuasive power of directive communication, which often shapes market sentiment more decisively than neutral or ambiguous statements (Tetlock et al., 2008).
The scenario simulation experiment further clarifies the psychological pathways through which opinion leaders operate. Their presence boosts investor confidence, reducing disagreements through reassurance, while enhancing cognitive processing by framing and contextualizing available information. These findings contribute to the growing literature on the intersection of social influence and financial decision-making (e.g., Cookson et al., 2024; Corgnet et al., 2024; Schmidt et al., 2020) by showing that opinion leaders function not only as opinion amplifiers but also as interpreters and validators of complex financial information.
Theoretical Contributions
This paper makes several theoretical contributions. Although some of our findings relate to domains outside OM, we believe these cross-domain insights enrich the OM perspective and constitute a distinct contribution of this paper. First, our findings refine information cascade theory in the context of financial social media. Contrary to the current OM literature that focuses on an information aggregation function (e.g., Ke et al., 2024; Xiao et al., 2021), our micro-level analysis reveals that opinion leaders do more than trigger cascades: they shape the quality of collective outcomes. Contrary to long-standing concerns that cascades suppress individualism and dilute decision quality (Wang et al., 2018), our findings show that cascades initiated by credible leaders can improve accuracy. This challenges the traditional view that cascades primarily reduce efficiency and suggests that they can function as mechanisms for filtering noise and consolidating useful information. Moreover, the moderating effects of order of appearance, expression style, and content novelty extend information cascade theory by showing that communication features influence not only the occurrence of cascades but also their effectiveness. These results raise an important question for future research: Under what conditions do cascades transition from harmful herd behavior to beneficial collective intelligence?
Second, we extend literature on financial service operations. Operational risk and decision efficiency in financial markets have long been central concerns in OM (Liu et al., 2023; Xu et al., 2017; Zhang et al., 2022). Studies typically focus on the technical progress, paying less attention to the prediction power of social media content. Although prior financial studies focus on extracting predictive signals from aggregated discussions (Aggarwal et al., 2012; Ben-Rephael et al., 2017; Chen et al., 2014), we demonstrate that contributions are not equal across participants. By showing that opinion leaders disproportionately improve prediction accuracy and accelerate price discovery, we highlight the need to reconsider the assumption of participant homogeneity in crowd-based models. This finding calls for the literature to account for hierarchies of influence within crowds and opens the door for theories of the wisdom of crowds to more explicitly integrate leader–follower dynamics.
Third, we enrich OM research on UGC platforms. Previous OM studies focus primarily on how to extract value from UGC (Cui et al., 2018; Ko et al., 2019; Yan et al., 2019; Yan and Pedraza-Martinez, 2019). More recent work extends this by examining content dynamics in sharing economy and crisis contexts (Li et al., 2024; Tang et al., 2024). In contrast, we highlight how OM factors, specifically the presence and role of opinion leaders, shape the content itself. Our findings complement those of Cookson et al. (2024), who emphasize how user type and platform design drive predictive power, by showing that cultivating credible opinion leaders enhances platform reputation and attracts sustained engagement. This suggests that platform managers should view opinion leaders not simply as participants but as strategic assets, raising further questions about how to balance leader influence with inclusivity and diversity of voices.
Fourth, we contribute to the literature on social media influencers by identifying independent opinion leaders as a distinct category. We acknowledge that research on opinion leaders remains relatively limited within the OM literature. This gap is notable given that opinion leaders increasingly shape information diffusion, user behavior, and market dynamics across digital platforms—phenomena that are central to many OM contexts. Expanding this literature can benefit operations managers by offering a deeper understanding of how influential actors affect the efficiency of decision-making. Unlike paid influencers (Chae et al., 2017) or organizational advocates (Ballantyne, 2000), independent leaders in financial forums derive their authority from demonstrated expertise and peer recognition rather than sponsorship. This finding suggests the need to broaden influencer research to account for credibility gained through community recognition rather than financial or organizational backing.
Managerial Implications
Our findings highlight the significant role that independent influencers play in shaping information flows and transactions on financial social media platforms. These insights benefit several stakeholders from an operational perspective. For individual investors, credible influencers can serve as valuable information intermediaries that help them interpret market signals and navigate increasingly complex financial content. By learning to distinguish trustworthy influencers from speculative voices, investors can make more informed decisions and reduce susceptibility to herd behavior.
For regulators, influencer-driven sentiment offers a real-time window into market dynamics. Monitoring these patterns can help detect emerging bubbles, misinformation, or coordinated behaviors that pose risks to market integrity. Insights into influencer effects can thus enhance regulatory oversight and early warning mechanisms.
For financial institutions, trusted influencers present opportunities to improve investor education and engagement. Collaborating with reputable influencers, particularly those who exhibit accuracy and responsible communication, may help institutions disseminate high-quality financial knowledge and better understand behavioral responses in retail investing.
Finally, for financial social media platforms, the findings underscore the importance of refining reputation systems and content-ranking mechanisms. Platforms can strengthen their governance by prioritizing credibility and information quality rather than popularity alone, thereby promoting healthier information environments and reducing the likelihood of harmful cascades.
Limitations and Future Research
Our study has several limitations that present opportunities for future research. First, this study is limited to scenarios involving a single opinion leader when assessing the moderating roles of order of appearance and expression style. This decision stems from the complexity inherent in situations involving multiple opinion leaders, who may post conflicting opinions in different styles. The dynamics of information cascades involving multiple opinion leaders presents a more complex scenario and will be explored in future research. Second, our empirical investigation focuses only on pre-COVID-19 data, which limits our ability to explore the temporal dynamics of opinion leaders’ influence, particularly in the postpandemic period. However, we believe that our empirical results should not exhibit significant changes before and after COVID-19. Our empirical results represent consistent human behavior that leverages fundamental psychological and social processes such as information cascades, herd behavior, and the wisdom of crowds (Ke et al., 2024; Tump et al., 2020). Third, although we find that opinion leaders cause information cascades, there may be heterogeneous effects for different investors. For example, an experienced investor may be less influenced by leaders’ opinions than an inexperienced investor. We plan to collect more information on individual investors (e.g., through surveys) and investigate these heterogeneous effects in the future. Fourth, we employed the BERT model to conduct sentiment analysis. It is important to note that both manual coding and algorithmic approaches may introduce measurement error, which can in turn add noise to the regression results. Future work may consider using more advanced techniques (e.g., large language models) to improve the accuracy of sentiment classification and enhance the robustness of empirical findings.
Supplemental Material
sj-docx-1-pao-10.1177_10591478261433261 - Supplemental material for The Role of Opinion Leaders in Crowd Wisdom: An Information Cascade Perspective
Supplemental material, sj-docx-1-pao-10.1177_10591478261433261 for The Role of Opinion Leaders in Crowd Wisdom: An Information Cascade Perspective by Shuo Yan, Kun Chen, Shaobo Li and Zhijie Lin in Production and Operations Management
Footnotes
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Natural Science Foundation of China [Grant numbers 72531006, 72472068, 72325002, 72421001, 72342009, and 72325010], Guangdong Basic and Applied Basic Research Foundation [Grant number 2025B1515020057], and Shenzhen Natural Science Foundation [Grant number JCYJ20250604144252069].
Notes
How to cite this article
Yan S, Chen K, Li S and Lin Z (2026) The Role of Opinion Leaders in Crowd Wisdom: An Information Cascade Perspective. Production and Operations Management 35(9): 3410–3428.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
